CALA-HSI: content-adaptive local attention for spectral-spatial hyperspectral image classification
Measurement Science and Technology · 07.08.2026 · 37 (31) · 315404
DOI: 10.1088/1361-6501/ae8d4a ↗Kurz gefasst
Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 05.08.2026. Verfasst von Mücahit CİHAN. Die Arbeit nennt 50 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Hyperspectral image (HSI) classification remains challenging because spectral bands are highly redundant, labeled samples are often scarce, and land-cover patterns can vary substantially across heterogeneous scenes. To address these challenges, this study proposes content-adaptive local attention for HSI classification (CALA-HSI), a compact spectral-spatial framework designed for robust classification under limited supervision. Instead of modeling all spatial positions globally, the proposed CALA mechanism adaptively aggregates spatial evidence within local neighborhoods. This allows each pixel to use context-dependent information while preserving a lightweight computational structure. A compact squeeze-and-excitation inspired spectral attention module (SpA) is also integrated before CALA to recalibrate informative spectral responses and suppress less useful channels. CALA-HSI is evaluated on four widely used benchmarks, Indian Pines, Pavia University, Salinas Scene, and Houston University, under four supervision regimes corresponding to 15%, 10%, 5%, and 1% labeled training samples. Across 16 evaluation scenarios and nine representative baselines, CALA-HSI achieves the highest mean overall accuracy (OA) of 96.04 ± 5.74%, the highest mean kappa coefficient of 94.28 ± 7.65%, the lowest mean loss of 0.153 ± 0.223, and the best mean OA rank of 1.44. It ranks first in OA and kappa coefficient in 11 out of 16 scenarios and appears in the Top-3 OA results in 15 scenarios. The model remains highly compact, with 46 724 parameters and an average computational cost of 0.0257 billion floating-point operations (GFLOPs). Ablation studies show that CALA is the main contributor to spectral-spatial discrimination, while the complete SpA and CALA design provides the most reliable gains, particularly under extreme label scarcity. Statistical and best/worst-case analyses further indicate a favorable scenario-level tendency for CALA-HSI. These results show that CALA-HSI provides an effective balance between accuracy, robustness, and efficiency for HSI classification under limited supervision.
Zitieren
APA
CİHAN, M. (2026). CALA-HSI: content-adaptive local attention for spectral-spatial hyperspectral image classification. Measurement Science and Technology, 37(31), 315404. https://doi.org/10.1088/1361-6501/ae8d4a
Vancouver
CİHAN M. CALA-HSI: content-adaptive local attention for spectral-spatial hyperspectral image classification. Measurement Science and Technology. 2026;37(31):315404. doi:10.1088/1361-6501/ae8d4a
BibTeX
@article{CHAN2026,
author = {CİHAN, Mücahit},
title = {CALA-HSI: content-adaptive local attention for spectral-spatial hyperspectral image classification},
journal = {Measurement Science and Technology},
year = {2026},
volume = {37},
number = {31},
pages = {315404},
doi = {10.1088/1361-6501/ae8d4a}
}
EUVIMED